Incremental learning method for object detection combining distillation learning and meta-learning
By combining distillation learning and meta-learning in a two-stage target detection model, the problems of forgetting old category knowledge and unstable prediction of new and old categories in incremental learning are solved, achieving efficient and stable detection of new and old categories and improving the flexibility and accuracy of the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- THE FIRST AFFILIATED HOSPITAL OF NAVAL MEDICAL UNIVERSITY OF CHINESE PEOPLES LIBERATION ARMY
- Filing Date
- 2022-07-13
- Publication Date
- 2026-07-21
Smart Images
Figure CN115311488B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer vision and relates to an object detection method, specifically an incremental learning method for object detection that combines distillation learning and meta-learning. Background Technology
[0002] Object detection technology, a core application technology in computer vision, aims to determine whether a class of interest exists in an input video or image, and, if so, to accurately determine the size and location of the target class. In recent years, with the continuous development of deep learning, especially the excellent performance of convolutional neural networks in pattern recognition, more and more object detection tasks have become highly automated and successfully applied in real-world scenarios.
[0003] However, in traditional machine learning and deep learning tasks, the categories of training data are fixed, and it is also assumed that the categories of test data are fixed. This leads to many models performing well on the test set but exhibiting poor transfer learning capabilities. This problem is even more pronounced in the practical application of object detection technology. The goal of incremental learning is to enable the system to quickly and continuously learn new knowledge from new samples while retaining most of the previously learned knowledge, similar to the human learning process. In recent years, with the continuous development of deep learning, incremental learning has been considered one of the main methods to significantly improve the transfer learning capabilities of models.
[0004] Current incremental learning is often based on distillation learning techniques. While distillation loss in distillation learning can successfully and quickly learn new knowledge, it also makes the learning of new knowledge in incremental learning highly dependent on the original dataset. Furthermore, as new knowledge is continuously learned, the model tends to memorize new knowledge, leading to a forgetting of older knowledge. This results in the network's improved accuracy on new categories at the cost of a significant decrease in accuracy on older categories. In real-world object detection tasks, high accuracy is often required for both new and old categories, making it difficult for the model to be used flexibly in real-world scenarios. Additionally, there is a problem of significant inconsistencies in prediction accuracy due to different training orders for new categories. This makes the model performance unstable and difficult to use in real-world scenarios. Summary of the Invention
[0005] To address the aforementioned problems, this invention provides a method capable of accurately detecting new categories with varying order of new categories and a small number of new category samples, while maintaining equally accurate detection of old categories. The technical solution adopted in this invention is as follows:
[0006] This invention provides an incremental learning method for target detection that combines distillation learning and meta-learning. The method comprises the following steps: Step S1, preprocessing the image to be tested; Step S2, constructing an incremental learning model for new and old target detection by combining distillation learning and meta-learning, inputting the preprocessed image to be tested into the model for detection to obtain the new and old category target detection results of the image. The incremental learning model for new and old target detection is based on a two-stage target detection module, including a first-stage foreground feature extraction sub-module based on distillation loss, and a second-stage classification and regression sub-module based on meta-learning technology. The foreground feature extraction sub-module includes, in sequence, an input layer, a general feature extractor backbone network module, a region proposal network module, and a first-stage output layer. The classification and regression sub-module includes, in sequence, an interest-aligned pooling layer, a regressor layer, a classifier layer, and a second-stage output layer.
[0007] The incremental learning method for target detection combining distillation learning and meta-learning provided by this invention may also have the following technical features, wherein the incremental learning models for new and old target detection are obtained through the following method: Step T1, the training set is divided into an old category training set and a new category training set; Step T2, an initial incremental learning model for old category target detection is constructed; Step T3, the old category training set is input into the initial incremental learning model for old category target detection to train and obtain an incremental learning model for old category target detection that performs target detection on the old category; Step T4, a new category learning branch is added to the incremental learning model for old category target detection and initialized to obtain an initial incremental learning model for new and old target detection; Step T5, the new category training set is input into the initial incremental learning model for new and old target detection to train until the training completion condition is met to obtain an incremental learning model for new and old target detection that can perform target detection on both the old and new categories simultaneously.
[0008] The incremental learning method for object detection combining distillation learning and meta-learning provided by this invention may also have the following technical features, wherein step T3 includes the following sub-steps: Step T3-1, given the initial old category object detection incremental learning model constructed in step T2; Step T3-2, input each training image in the old category training set into the initial old category object detection incremental learning model in sequence and perform one iteration; Step T3-3, calculate the loss error using the model parameters of the last layer and backpropagate the loss error to update the model parameters; Step T3-4, repeat steps T3-2 to T3-3 until the training completion condition is met, and obtain the trained old category object detection incremental learning model.
[0009] The incremental learning method for object detection combining distillation learning and meta-learning provided by this invention may also have the following technical features, wherein step T5 includes the following sub-steps: Step T5-1, given the initial new and old incremental learning models for object detection constructed in step T4; Step T5-2, based on the new category training set, sequentially acquire a batch of new category training images and input them into the initial new and old incremental learning models for one iteration; Step T5-3, use the model parameters of the last layer to calculate the loss errors of the new category and the old category respectively, and backpropagate the loss errors to their respective category branches to update the model parameters; Step T5-4, repeat steps T5-2 to T5-3 until the training completion condition is met, and obtain the new and old incremental learning models for object detection.
[0010] The incremental learning method for target detection combining distillation learning and meta-learning provided by this invention may also have the following technical features: the new and old target detection incremental learning models are constructed based on the new category learning branch added to the old category target detection incremental learning model. During the training process of the new and old target detection incremental learning models, the two-stage target detection module adopts a segmented incremental learning strategy for incremental learning, as follows: Step P1, the image data in the training set is input into the new and old target detection incremental learning models. The old category target detection incremental learning model and the new category learning branch in the model obtain the foreground prediction results covering the old category and the new category respectively based on the distillation loss for the foreground feature extraction submodule. The loss is calculated based on the output results and backpropagation is performed to train the foreground feature special area submodule. Step P2, the foreground prediction results covering the new and old categories are used as the input of the second-stage classification and regression submodule based on meta-learning technology to obtain the accurate prediction results of different categories of targets. The loss is calculated based on the output results and backpropagation is performed to train the classification and regression submodule.
[0011] The incremental learning method for object detection combining distillation learning and meta-learning provided by this invention may also have the following technical features, wherein the training process of the first-stage foreground feature extraction submodule based on distillation loss is as follows: Step P1-1, obtain the foreground prediction result of the old category corresponding to the input image data based on the old category object detection incremental learning model; Step P1-2, regard the foreground prediction result of the old category as the true label of the old category foreground box; Step P1-3, input each training image in the training set into the new category learning branch in sequence, and perform one iteration to obtain the foreground prediction result belonging to the old and new categories; Step P1-4, separate the foreground prediction results belonging to the old and new categories according to the old and new categories; Step P1-5, apply the classification loss error to the new category and the distillation loss to the old category, and then perform a weighted sum of the two loss errors to obtain the final loss error; Step P1-6, backpropagate the final loss error to update the parameters of the new and old object detection incremental learning model.
[0012] The incremental learning method for target detection combining distillation learning and meta-learning provided by this invention may also have the following technical features, wherein the training process of the classification and regression sub-module based on meta-learning technology in the second stage is as follows: Step P2-1, the foreground prediction results obtained in step P1 are sequentially input into the classification and regression sub-module based on meta-learning technology in the second stage for one iteration to obtain the prediction results of the classifier and regressor belonging to the new category and the old category, respectively; Step P2-2, the classification and regression loss errors are obtained based on the prediction results of the classifier and regressor and the true label values, respectively, and then the two loss errors are weighted and summed to obtain the final loss error; Step P2-3, the final loss error is backpropagated to update the parameters of the incremental learning model for target detection of the new and old objects.
[0013] Invention Function and Effect
[0014] The incremental learning method for object detection combining distillation learning and meta-learning according to the present invention decouples the classification processes of different categories in the original model using meta-learning, enabling the model to initially detect new categories. Furthermore, decoupling different categories isolates the learning process of new knowledge from old knowledge, significantly reducing the impact on the detection performance of old categories. Further, the meta-learning approach allows the optimization process of the new category classifier parameters to be shared with the old category classifier, enabling the new category classifier to converge quickly, thus facilitating incremental learning of new categories with fewer samples. Simultaneously, leveraging the "segmented" detection property of the two-stage object detection model, distillation learning enhances the feature extractor's ability to extract features of new categories, thereby further improving the foreground detection performance of new categories with minimal impact on the foreground detection performance of old categories.
[0015] Therefore, the incremental learning method for target detection that combines distillation learning and meta-learning provided by this invention can simultaneously and efficiently detect new categories of targets added in different orders and existing old categories of targets in real complex scenes through the incremental learning model for target detection, and has the advantages of strong generalization and high accuracy. Attached Figure Description
[0016] Figure 1 This is a flowchart illustrating the incremental learning method for target detection that combines distillation learning and meta-learning in an embodiment of the present invention.
[0017] Figure 2 This is a schematic diagram of the structure of the incremental learning model for new and old target detection in an embodiment of the present invention;
[0018] Figure 3 This is a schematic diagram of the training process of the incremental learning model for new and old target detection in an embodiment of the present invention;
[0019] Figure 4 This is a schematic diagram illustrating the process of segmented incremental learning between the old and new target detection incremental learning models in an embodiment of the present invention. Detailed Implementation
[0020] To make the technical means, creative features, objectives and effects of this invention easier to understand, the following describes the incremental learning method for target detection that combines distillation learning and meta-learning in conjunction with the embodiments and accompanying drawings.
[0021] <Example>
[0022] In this embodiment, the incremental learning method for object detection combining distillation learning and meta-learning is run on a single computer and trained using eight graphics cards. Each pair of graphics cards forms a group for iterative training. In this group, the first card performs forward propagation calculations for the old class network model, and the second card performs forward and backward propagation calculations for the new class network model. The trained model using the incremental learning method for object detection combining distillation learning and meta-learning is stored in the computer as executable code.
[0023] Figure 1 This is a flowchart illustrating the incremental learning method for target detection that combines distillation learning and meta-learning in an embodiment of the present invention.
[0024] like Figure 1 As shown, the incremental learning method for target detection that combines distillation learning and meta-learning includes the following steps:
[0025] Step S1: Preprocess the image to be tested.
[0026] In this embodiment, preprocessing includes image filling, image normalization, random image flipping, rotation, and occlusion operations at different positions of the target.
[0027] Step S2: Combine distillation learning and meta-learning to construct an incremental learning model for new and old object detection. Input the preprocessed image to be tested into the model for detection to obtain the new and old category object detection results of the image to be tested.
[0028] Figure 2 This is a schematic diagram of the structure of the incremental learning model for new and old target detection in an embodiment of the present invention.
[0029] like Figure 2 As shown, the incremental learning model for object detection, combining the old and new models, is built upon a two-stage object detection module. This module includes a first-stage foreground feature extraction submodule based on distillation loss, and a second-stage classification and regression submodule based on meta-learning techniques. The foreground feature extraction submodule comprises, in sequence, an input layer, a general feature extractor backbone network module, a region proposal network module, and a first-stage output layer. The classification and regression submodule comprises, in sequence, an interest-aligned pooling layer, a regressor layer, a classifier layer, and a second-stage output layer.
[0030] Specifically, the regressor layer is set to a "class-independent" regressor module, and the classifier layer is set to a "class-specific" classifier module.
[0031] The "category-specific" classifier module contains a "category-specific" representation generator submodule. The input to the representation generator is local features F, and the representation generator submodule consists of a simple 3x3 convolutional layer, a pooling layer, and another 3x3 convolutional layer. The generated representation gradient is represented by C. k This is indicated. Then, the "class-independent" regressor module uses C to represent the local features F and the characterization gradient. k Batch Normalization is performed on each sample. Then, the processed F' and C are... k Performing dot product operation yields preliminary classification result Y. k Where F' has dimensions {1, N, C1}, C k The dimension of ' is {1, C1, K}, Y k The dimensions are {1, N, K}, where N is the number of candidate boxes in the current image, C1 is the number of feature points, and K is the number of currently detectable categories. Finally, the Sigmoid function is used to transform Y... k Further calculations yielded the final classification result.
[0032] The “class-independent” regressor module integrates features using a 3x3 convolution, and then uses two 1x1 convolutions to obtain the position offset of the predicted bounding box.
[0033] In this embodiment, the output dimension of the incremental learning model for new and old object detection is eighty-five dimensions. The first eighty dimensions represent the probability scores of the current predicted box being identified as one of the eighty categories in the image, the eighty-first dimension represents the probability score of the current predicted box being identified as the background, and the last four dimensions represent the positional offset of the current predicted box in the image, namely the x and y offsets of the center point and the length and width offsets.
[0034] In this embodiment, the new and old target detection incremental learning model that combines distillation learning and meta-learning is first trained through the model training step and stored in the computer. The computer can call the model through executable code and simultaneously process the images to be detected in batches to obtain and output the different category classification and location prediction results for each image to be detected.
[0035] Figure 3 This is a schematic diagram of the training process of the incremental learning model for new and old target detection in an embodiment of the present invention.
[0036] like Figure 3 As shown, its training process is as follows:
[0037] Step T1: Divide the training set into an old category training set and a new category training set.
[0038] In this embodiment, the dataset used is the challenging publicly available object detection dataset MS-COCO2017. It contains 115,000 images for training and 5,000 images for validation, covering 80 object classes, 20 of which are precisely the 20 classes covered by another publicly available dataset, PASCAL VOC. Therefore, these 20 object classes are considered new classes, and are currently referred to as the augmenting classes. The remaining 60 classes are considered old classes, and are currently referred to as the base classes. In this embodiment, the base classes are used as the old class training set. During each training process for learning branches for new classes, three augmenting classes are sequentially sampled as the new class training set.
[0039] Step T2: Construct the initial old category object detection incremental learning model.
[0040] In this embodiment, the incremental learning model for old category object detection contains different model parameters in each layer, and these model parameters are randomly set during construction.
[0041] In this embodiment, the initial incremental learning model for old-category object detection uses the two-stage object detection model Faster R-CNN as the baseline model. In the first stage of the detector, the ResNet50 residual network model is used as the feature extraction backbone network model, and the model is built using the existing deep learning framework PyTorch. In the second stage of the detector, a classifier and regressor composed of convolutional layers, pooling layers, activation layers, and batch normalization layers are used.
[0042] Step T3: Input the old category training set into the initial old category object detection incremental learning model to train and obtain an old category object detection incremental learning model that performs object detection on the old category.
[0043] In step T3 of this embodiment, images from the training set are fed into the network model in batches for training. The batch size of training images fed into the network model each time is 16, and the training is iterated for a total of 35 rounds. Step T3 includes the following sub-steps:
[0044] Step T3-1: Given the initial old category object detection incremental learning model constructed in step T2;
[0045] Step T3-2: Input each training image from the old category training set into the initial old category object detection incremental learning model in sequence and perform one iteration;
[0046] Step T3-3: Calculate the loss error using the model parameters of the last layer and backpropagate the loss error to update the model parameters;
[0047] Step T3-4: Repeat steps T3-2 to T3-3 until the training completion condition is met, and obtain the old category object detection incremental learning model after training.
[0048] In the model training process described above, after each iteration (i.e., after the training image passes through the model), the loss error of the model parameters in the last layer of both stages of the incremental learning model for object detection in this embodiment is calculated, and then the calculated loss error is backpropagated. In the training of the old category object detection incremental learning model, the regression loss in both the first and second stages adopts the Smooth l1 loss function, the classification loss in the first stage adopts the cross-entropy loss, and the classification loss in the second stage adopts the Sigmoid loss. During the training process, the learning rate is optimized according to the "warm-up" method to update the model parameters. The learning rate is initialized to 0.001, and the learning rate is increased linearly in the first five iterations to search for the optimal path. After the fifth iteration, the learning rate is increased to 0.01. After that, the learning rate is decreased, and after the 11th iteration, the learning rate is reduced by 10%. In addition, the training completion conditions of the model are the same as those of the conventional convolutional neural network model, that is, the training is completed after the model parameters of each layer converge.
[0049] Step T4: Add a new category learning branch to the old category object detection incremental learning model and initialize it to obtain the initial new and old object detection incremental learning model.
[0050] The incremental model for object detection in this embodiment is obtained by adding some modules specific to the new category to the incremental learning model for object detection in the old category. Each layer contains different model parameters. Among them, the parameters of the existing layers of the incremental learning model for object detection in the old category are consistent with those of the old category during retraining. However, the parameters of the newly added layers are randomly set during construction.
[0051] Step T5: Input the new category training set into the initial new and old object detection incremental learning model for training until the training completion condition is met, resulting in a new and old object detection incremental learning model that can perform new and old category object detection simultaneously.
[0052] In step T5 of this embodiment, images from the new category training set are fed into the network model in batches for training. The batch size of training images entering the network model each time is 16, and the training is iterated for a total of 10 rounds. Step T4 includes the following sub-steps:
[0053] Step T5-1: Given the initial incremental learning model for new and old object detection constructed in step T4;
[0054] Step T5-2: Based on the new category training set, a batch of new category training images are sequentially obtained and input into the initial new and old object detection incremental learning model for one iteration;
[0055] Step T5-3: Calculate the loss errors of the new and old classes using the model parameters of the last layer, and backpropagate the loss errors to their respective class branches to update the model parameters.
[0056] Step T5-4: Repeat steps T5-2 to T5-3 until the training completion condition is met, and obtain the new and old object detection incremental learning models after training.
[0057] In the model training process described above, after each iteration (i.e., after the training image passes through the model), the loss error of the model parameters in the last layer of both stages of the target model in this paper is calculated. Then, the calculated loss error is backpropagated, and the stochastic gradient descent algorithm is used to optimize the parameters, thereby updating the model parameters. In the initial incremental learning model training for new and old object detection, the old category loss in the first stage uses distillation loss; in this experiment, the distillation loss uses the Smooth L1 loss function. For the new category classification loss, cross-entropy loss is used, and the regression loss uses the Smooth L1 loss function. In the second stage, the regression loss for new and old categories uses the Smooth L1 loss function. The classification loss for new and old categories in the second stage uses the Sigmoid loss. During training, the learning rate is initialized to 0.001, and the stochastic gradient descent algorithm is used to optimize the parameters, thereby updating the model parameters. Furthermore, the training completion conditions are the same as for conventional convolutional neural network models; that is, training is complete once the model parameters of each layer converge.
[0058] In this embodiment, after training is completed in step T5, the obtained incremental learning model for object detection is validated on the validation set. The experimental results show that for images to be detected containing eighty classes, the average accuracy can reach 36.3% on the sixty base classes and 26.9% on the twenty new classes.
[0059] In this embodiment, the incremental learning model for new and old object detection, which combines distillation learning and meta-learning, employs a segmented incremental learning strategy in its two-stage object detection module for incremental learning of new categories. Specifically:
[0060] Figure 4 This is a schematic diagram illustrating the process of segmented incremental learning between the old and new target detection incremental learning models in an embodiment of the present invention.
[0061] In this embodiment, each image to be detected is simultaneously input into both the old and new models. To prevent storing too many model parameters on a single graphics card and thus reducing model convergence performance, this embodiment uses eight NVIDIA 1080Ti graphics cards for training, with each pair of graphics cards forming a group for iterative training. In this group of two graphics cards, the first card performs forward propagation calculations for the old class network model, and the second card performs forward and backward propagation calculations for the new class network model. Figure 4 As shown, the specific process is as follows:
[0062] Step P1: Input the image data in the training set into the incremental learning model for new and old object detection. The old category object detection incremental learning model and the new category learning branch in the model, based on the foreground feature extraction submodule with distillation loss, obtain the foreground prediction results covering the old category and the new category respectively. The loss is calculated based on the output results and backpropagation is performed to train the foreground feature special area submodule.
[0063] In this step, the training process for the foreground feature extraction submodule based on distillation loss in the first stage is as follows:
[0064] Step P1-1: Obtain the foreground prediction results of the old category corresponding to the input image data based on the old category object detection incremental learning model;
[0065] Step P1-2: Treat the foreground prediction results of the old category as the true labels of the foreground boxes of the old category;
[0066] Steps P1-3: Input each training image in the training set into the new category learning branch in sequence, and perform one iteration to obtain the foreground prediction results belonging to the old and new categories.
[0067] Steps P1-4 separate the prospect prediction results belonging to the old and new categories according to the old and new categories;
[0068] Steps P1-5: Apply classification loss error to the new category and distillation loss to the old category, then weight and sum the two loss errors to obtain the final loss error.
[0069] Steps P1-6 involve backpropagating the final loss error to update the parameters of the old and new incremental learning models for object detection.
[0070] Step P2 takes the foreground prediction results covering both new and old categories as input to the classification and regression submodule based on meta-learning technology in the second stage, obtains accurate prediction results for different categories of targets, and calculates the loss based on the output results and backpropagates to train the classification and regression submodule.
[0071] In this step, the training process of the classification and regression submodule based on meta-learning technology in the second stage is as follows:
[0072] Step P2-1: Input the foreground prediction results obtained in step P1 into the classification and regression submodule based on meta-learning technology in the second stage, perform one iteration, and obtain the classifier and regressor prediction results belonging to the new category and the old category, respectively.
[0073] Step P2-2: Based on the prediction results of the classifier and regressor and the true label values, obtain the classification and regression loss errors respectively, and then sum the two loss errors by weight to obtain the final loss error.
[0074] Steps P2-3 involve backpropagating the final loss error to update the parameters of the old and new object detection incremental learning models.
[0075] In this embodiment, to address the challenge of cross-GPU data communication when performing a task on two GPUs, the NVIDIA Multi-GPU Communication Framework (NCCL) integrated into the GPU is utilized to implement three cross-GPU data transmissions. These three transmissions are as follows: The first transmission occurs at the output of the first-stage candidate box region proposal network, where the output of the old class network is transmitted to the new class learning branch. The second transmission transmits the candidate box ROIs output from the first-stage candidate box region proposal network of the new class learning branch to the old class network, allowing it to perform the second-stage training based on the ROIs predicted by the new class. The third transmission occurs at the output of the second-stage candidate box region proposal network, again requiring the transmission of the old class network's output to the new class learning branch.
[0076] In this embodiment, the incremental learning model for target detection, which combines distillation learning and meta-learning, learns new class knowledge without affecting the detection performance of the old class, and performs target detection on the image based on this feature.
[0077] Functions and effects of the embodiments
[0078] The incremental learning method for object detection combining distillation learning and meta-learning provided in this embodiment constructs an incremental learning model for both new and old object detection based on a two-stage object detection module. This model includes a first-stage foreground feature extraction submodule based on distillation loss and a second-stage classification and regression submodule based on meta-learning techniques. This approach enables accurate detection of new categories with different orderings and a small number of new category samples, while maintaining equally accurate detection of old categories. By fully utilizing meta-learning techniques to provide different learning gradients for different categories to prevent interference between categories, and by employing distillation loss to enhance the feature extraction module, more accurate object detection for both new and old categories is achieved. This not only avoids significantly impacting the knowledge learned by the old category training model but also promotes rapid knowledge learning for the new category with a small number of training samples.
[0079] In this embodiment, the incremental learning model for object detection uses the Faster R-CNN object detection model as the baseline model, and optimizes the two-stage models using distillation loss and meta-learning techniques respectively. Therefore, the model not only possesses the advantage of distillation loss in quickly transferring new categories to the model, but also successfully avoids the problems of "catastrophic forgetting" and "sequential interference" that often occur with distillation loss methods. Furthermore, the meta-learning-based incremental learning method not only makes the model more flexible in adding new categories, but also makes learning with few samples more convenient. This embodiment achieves good performance on the publicly available MS-COCO dataset in a complex environment, demonstrating the effectiveness of the proposed method.
[0080] The above embodiments are only used to illustrate specific implementations of the present invention, and the present invention is not limited to the scope of the description of the above embodiments.
Claims
1. A target detection incremental learning method combining distillation learning and meta-learning, used to achieve accurate new category detection on a small number of new category samples with different new category orders, while ensuring equally accurate old category detection, characterized in that... Includes the following steps: Step S1: Preprocess the image to be tested; Step S2: Combine distillation learning and meta-learning to construct an incremental learning model for new and old object detection. Input the preprocessed image to be tested into this model for detection to obtain the new and old category object detection results of the image to be tested. The incremental learning model for object detection, which combines old and new models, is built upon a two-stage object detection module. This module includes a first-stage foreground feature extraction sub-module based on distillation loss, and a second-stage classification and regression sub-module based on meta-learning techniques. The foreground feature extraction submodule includes, in sequence, an input layer, a general feature extractor backbone network module, a region proposal network module, and a first-stage output layer. The classification and regression submodule includes an interest-aligned pooling layer, a regressor layer, a classifier layer, and a second-stage output layer, arranged sequentially. The incremental learning model for object detection, which combines the old and new models, is obtained through the following method: Step T1: Divide the training set into an old category training set and a new category training set; Step T2: Construct the initial old category object detection incremental learning model; Step T3: Input the old category training set into the initial old category object detection incremental learning model to train and obtain an old category object detection incremental learning model for object detection of the old category; Step T4: Add a new category learning branch to the old category object detection incremental learning model and initialize it to obtain the initial new and old object detection incremental learning model; Step T5: Input the new category training set into the initial incremental learning model for new and old object detection for training until the training completion condition is met, thus obtaining an incremental learning model for new and old object detection that simultaneously performs new and old category object detection. The new and old object detection incremental learning model is constructed based on the new category learning branch added to the old category object detection incremental learning model. During the training process of this new and old object detection incremental learning model, its two-stage object detection module adopts a segmented incremental learning strategy to perform incremental learning of the new category, as follows: Step P1: Input the image data in the training set into the incremental learning model for new and old object detection. The old category object detection incremental learning model and the new category learning branch in the model, based on the distillation loss, the foreground feature extraction submodule obtains the foreground prediction results covering the old category and the new category respectively. The loss is calculated based on the output results and backpropagation is performed to train the foreground feature extraction submodule. Step P2 involves using the foreground prediction results covering both new and old categories as input to the classification and regression submodule based on meta-learning technology in the second stage. This yields accurate prediction results for different target categories. The loss is then calculated based on the output results, and backpropagation is performed to train the classification and regression submodule. The training process for the first-stage foreground feature extraction submodule based on distillation loss is as follows: Step P1-1: Based on the old category object detection incremental learning model, obtain the foreground prediction result of the old category corresponding to the input image data; Step P1-2: Treat the foreground prediction result of the old category as the true label of the foreground box of the old category; Steps P1-3: Input each training image in the training set into the new category learning branch in sequence, and perform one iteration to obtain the foreground prediction results belonging to the old and new categories; Steps P1-4 separate the prospect prediction results belonging to the old and new categories according to the old and new categories; Steps P1-5: Apply classification loss error to the new category and distillation loss to the old category, then weight and sum the two loss errors to obtain the final loss error. Steps P1-6 involve backpropagating the final loss error to update the parameters of the new and old incremental learning models for object detection.
2. The incremental learning method for target detection combining distillation learning and meta-learning according to claim 1, characterized in that: in, Step T3 includes the following sub-steps: Step T3-1: Given the initial old category object detection incremental learning model constructed in step T2; Step T3-2: Input each training image in the old category training set into the initial old category object detection incremental learning model in sequence and perform one iteration; Step T3-3: Calculate the loss error using the model parameters of the last layer and backpropagate the loss error to update the model parameters; Step T3-4: Repeat steps T3-2 to T3-3 until the training completion condition is met, and obtain the trained old category object detection incremental learning model.
3. The incremental learning method for target detection combining distillation learning and meta-learning as described in claim 1, Its features are: Step T5 includes the following sub-steps: Step T5-1: Given the initial incremental learning model for new and old object detection constructed in step T4; Step T5-2: Based on the new category training set, a batch of new category training images are sequentially obtained and input into the initial new and old target detection incremental learning model for one iteration; Step T5-3: Calculate the loss error of the new class and the old class using the model parameters of the last layer, and backpropagate the loss error to the respective class branches to update the model parameters. Step T5-4: Repeat steps T5-2 to T5-3 until the training completion condition is met, and obtain the new and old target detection incremental learning model.
4. The incremental learning method for target detection combining distillation learning and meta-learning as described in claim 1, Its features are: The training process of the classification and regression submodule based on meta-learning technology in the second stage is as follows: Step P2-1: Input the foreground prediction results obtained in step P1 into the classification and regression submodule based on meta-learning technology in the second stage in sequence, perform one iteration, and obtain the classifier and regressor prediction results belonging to the new category and the old category respectively. Step P2-2: Based on the prediction results of the classifier and regressor and the true label values, obtain the classification and regression loss errors respectively, and then sum the two loss errors by weight to obtain the final loss error. Step P2-3 involves backpropagating the final loss error to update the parameters of the new and old incremental learning models for object detection.